Development of Predictive Model based on Medical Image for Prediction of Diagnosis and Treatment Prognosis in Patients with Pulmonary Hypertension

Abstract

Pulmonary hypertension (PH) is a condition characterized by elevated pressure in the pulmonary arteries resulting from various pulmonary and cardiac diseases. This elevation in pressure increases the strain on the heart, leading to right ventricular overload and potentially right heart failure, which significantly affects the patient's prognosis. This study aims to develop a technology that supports non-invasive examinations by deriving hemodynamic computational fluid dynamics (CFD) simulation results from echocardiography and computed tomography (CT) images. These results will be used to develop a deep learning model (DL-CFD). Additionally, a model incorporating the clinical results of echocardiography (eDL-CFD) will be evaluated for validity. By comparing these models with traditional echocardiography methods, the study confirms the potential for clinical use of this combined approach. Between 2008 and 2019, a retrospective study analyzed 92 patients who underwent right heart catheterization (RHC) for PH assessment, including 75 diagnosed with PH (mPAP > 25 mmHg) and 17 suspected but not diagnosed (mPAP < 25 mmHg). The results of this study demonstrated the efficacy of integrating DL-CFD and eDL-CFD for non-invasive diagnosis of PH. The deep learning model's predictions closely matched the simulation results, showcasing high accuracy and reliability. Comparative analysis revealed that while the correlation between RHC and the developed CFD and DL-CFD methods was lower than traditional echocardiography, eDL-CFD method exhibited a higher correlation with RHC and improved diagnostic accuracy. The area under the curve (AUC) for the combined method was 98.9%, significantly higher than the 94.6% for echocardiography alone. Stratified analysis highlighted that the combined approach improved specificity to 94.1% from 76.4%, maintaining a high sensitivity of 97.3%. This indicates the potential of the combined method to serve as a more reliable non-invasive diagnostic tool for pulmonary hypertension. These results indicate that the eDL-CFD approach can be a viable non-invasive alternative for diagnosing PH, potentially offering improved diagnostic accuracy and reliability over traditional echocardiography alone. This study demonstrates the potential of eDL-CFD for non-invasive PH diagnosis, supporting personalized treatment planning and accurate prediction of disease progression. Future research will focus on utilizing diverse datasets and applying data augmentation techniques to enhance the model's generalizability and accuracy.

폐고혈압(pulmonary hypertension, PH)은 폐동맥 압력이 비정상적으로 상승하여 심장과 폐에 과도한 부담을 주며, 치료하지 않으면 심부전 등의 심각한 건강 문제를 초래할 수 있는 질환이다. 일반적으로 폐고혈압 검사는 우심도자술(right heart catheterization, RHC)이 표준 진단법(gold standard)이나, 침습검사로 환자에게 부담이 있을 수 있다. 반면, 비침습검사인 심초음파는 초기 스크리닝 도구로 유용하지만, 진단 정확도가 우심도자술에 비해 낮은 상황이다. 따라서 본 연구는 심초음파(echocardiography)와 컴퓨터 단층촬영(computed tomography, CT)으로 혈역학적 유체 역학(computational fluid dynamics, CFD) 시뮬레이션 결과를 도출하고, 이러한 결과를 이용해 딥러닝 모델(deep learning with CFD simulation, DL-CFD) 및 심초음파의 임상 결과를 포함한 모델(echocardiography + DL-CFD, eDL-CFD)을 개발하고 유효성을 평가하였다. 2008년부터 2019년까지, PH 평가를 위해 우심도자술(RHC)을 받은 92명의 환자를 대상으로 후향적 연구를 수행하였다. 이 중 75명은 PH로 진단되었고(평균 폐동맥압; mean pulmonary arterial pressure, mPAP > 25 mmHg), 17명은 정상군이다 (mPAP < 25 mmHg). 비교 분석 결과, RHC와 개발된 CFD 및 DL-CFD 방법 간의 상관관계는 전통적인 심초음파보다 낮았으나, eDL-CFD 방법은 RHC와 더 높은 상관관계를 보였다. eDL-CFD 방법의 AUC는 98.9%로, 심초음파 단독(94.6%)보다 높은 결과를 보였고, 특이도가 76.4%에서 94.1%로 향상되었으며, 민감도는 97.3%로 유지되었다. 이 결과 eDL-CFD 접근법이 심초음파 단독검사 보다 개선된 진단 정확도와 신뢰성을 제공할 수 있는 유효한 비침습적 대안임을 보였다. 또한 본 연구는 비침습적인 방법으로 개인 맞춤형 치료 계획 및 질병 진행에 대한 예측 가능성을 확인하였다. 향후 연구는 다양한 데이터셋을 활용하고 데이터 증강 기법을 적용하여 모델의 일반화 가능성과 정확성을 더욱 향상시키는 데 초점을 맞출 것이다.